Overview.¶
This project's main objective was to develop a machine learning solution for determining if a patient will be admitted to the intensive care unit. This is done with the hopes that ICU resources can be set up, patient transfers can be planned, frontline doctors may safely discharge these patients, and remote follow-up can be done with them.
The project was concerned with Covid-19 cases, with more than 16 million confirmed illnesses and 454,429 confirmed deaths by May 26, 2021, Brazil is one of the nations most impacted by the COVID-19 pandemic (according to the Johns Hopkins Coronavirus Resource Center). Brazil was one of the nations most affected by the first wave of Covid-19, which had its first case on February 26, 2020 and began spreading in communities on March 20, 2020, leaving Brazil unprepared and unable to respond because of the pressure on hospital capacity, including the lengthy and intense requests for ICU (incentive care unit) beds, staff, personal protection equipment, and medical resources.
Data¶
This dataset, which comprises of 1925 rows and 231 columns, was collected using anonymized data from the Hospital Sri-Libanês, So Paulo, and Brasilia. PATIENT VISIT IDENTIFIER, which includes 385 patients with various 5 rows of records, is the dataset's unique identifier. The target is the ICU columns, which indicates whether the patient was admitted or not using a binary representation of 1 for admission and 0 for non-admission, and the Window column indicate the time-period the patient was admitted. Other notable features in this dataset
- Patient demographic information (03)
- Patient previous grouped diseases (09)
- Blood results (36)
- Vital signs (06)
- Missing data was represented has Nan, consists of 223863 missing cells (50.3%), and no duplicated data
It was observed that 195 patients were admitted to ICU and 190 patients were not admitted
Data Cleaning and pre-processing.¶
The dataset was transpose to convert the unique identifier to one row per patient. Hence, converting the data to 385 rows and 1151 columns
Taking care missing values¶
We have 223,863 missing cells (50.3%), and some features having over 89% missing data Columns with over 50% overrepresentation of null values were dropped. The other missing values were imputed with backward fill as suggested by the dataset providers. This downsized the dataset to 230 columns and 394 rows, as one of the patient visit identifier has a complete missing values in the entire row. As instructed by the dataset providers, that the data gathered after the ICU admission should not be taken into consideration, and patient admitted in window_1 (0-2Hours). Hence, our dataset was reduced further to 352rows and 50 columns after eliminating this rows and columns During that exploration the dataset was divided into patient-constant features (patient constant features are features that contain the same value for a single patient across all time points) and time-variant features (Time-variant features are features that contain multiple values for the same patient, such as multiple lab test results for a single patient over time) for better visualization According to the patient-constant-features, more males and patients over the age of 65 were admitted to the ICU, while the time-variant-features were able to show correlation with the DIFF cluster and DIFF_REL.
Correlation Checking the correlation with visualization wasn’t helpful because of large data, the data was split into two, first we are looked at how these features relate to each other, excluding the target column, and a stacked format was adopted to better understand the correlation
Secondly, we looked at how these features relate to ICU
Feature Encoding Feature encoding was performed on the Age percentile has it consist of columns with object.
Methodology(Modelling and Model Evaluation)¶
The dataset was split into train and validation, 90% of the data was allocated to the train data and 10% to the validation. Ensemble Learning Methods was implemented using 8 modelling algorith, some of the results are particularly outstanding, especially given that 46% of genuine values in our desired frequency. A perfect model will have an ROC-AUC score of 1, while a model that is no better than random guessing will have an ROC-AUC score of 0.5. Hence, algorithm such as KNN, Decision Tree, SVM have low ROC-AUC score The results are significantly better on Random-Forest. This is an excellent sign because some of the models wasn't significantly more accurate than picking patients at random to be admitted to the intensive care unit Cross-validation has several advantages over other methods for evaluating the performance of a model. For example, it can provide a more accurate estimate of the model's performance because it uses more of the data for training and evaluation. It can also be used to tune hyper-parameters, which are model-specific parameters that cannot be learned from the data.
Hyper-parameters.
Feature selection and hyper-parameters was performed on the Random-Forest model to better enhanced the accuracy
We were able to enhance the validation accuracy by around 6% just by adjusting the algorithm's settings. Additionally, we have reached a point where our model should be able to forecast patients who will need an ICU bed in more than 80% of situations.
Conclusion and limitation¶
We were able to develop prediction models on this notebook for the ICU admission classification issue. For each patient, it concentrated on the earliest data available, creating a model that was reasonably accurate. The model's ability to successfully categorise patients for both goal values is one sign that this data processing stage was successful.
We must reiterate our caution that working with tiny datasets restricts how confident we can be in our findings.
